Implementation:NVIDIA DALI Fn Resize Linear

Knowledge Sources	NVIDIA DALI
Domains	Image_Processing, GPU_Computing, Image_Transformation
Last Updated	2026-02-08 00:00 GMT

Overview

Concrete operator for resizing image tensors to a target spatial resolution using configurable interpolation, provided by the nvidia.dali.fn module.

Description

fn.resize resamples image tensors from their original [H, W, C] dimensions to a specified target (height, width) using a configurable interpolation kernel. When operating on GPU-resident tensors (the typical case after mixed-device decoding), the entire resize computation runs on the GPU with no host-device data transfer.

Key behaviors:

• size accepts a (height, width) tuple specifying the target output dimensions.
• interp_type=types.INTERP_LINEAR selects bilinear interpolation, computing each output pixel as a weighted average of the four nearest source pixels.
• antialias=False disables the low-pass pre-filter that prevents aliasing during downsampling, trading visual quality for speed.
• The operator preserves the input data type (uint8 in, uint8 out) and channel count.
• Both upsampling and downsampling are supported within the same batch; each sample is resized independently.

Usage

Place fn.resize after the decode operator and before normalization or augmentation operators. Specify the target dimensions via the size parameter as a (height, width) tuple. Select the interpolation method via interp_type based on the quality/speed trade-off required by the application.

Code Reference

Source Location

• Repository: NVIDIA DALI
• File: docs/examples/zoo/images/decode_and_transform_pytorch.py (lines 85-90)

Signature

fn.resize(
    decoded,
    size=img_hw,
    interp_type=types.INTERP_LINEAR,
    antialias=False,
)

Import

import nvidia.dali.fn as fn
import nvidia.dali.types as types
# or
from nvidia.dali import fn, types

I/O Contract

Inputs

Name	Type	Required	Description
decoded	DataNode	Yes	Input image tensor with layout [H, W, C], typically GPU-resident after mixed-device decoding
size	tuple(int, int) or list	Yes	Target output dimensions as (height, width)
interp_type	types.DALIInterpType	No	Interpolation method: types.INTERP_LINEAR (bilinear), types.INTERP_NN (nearest neighbor), types.INTERP_CUBIC, types.INTERP_LANCZOS3, types.INTERP_TRIANGULAR. Default: types.INTERP_LINEAR
antialias	bool	No	Enable low-pass pre-filter to reduce aliasing when downsampling. Default: True

Outputs

Name	Type	Description
resized	DataNode (GPU)	Resized image tensor with layout [target_h, target_w, C] and same dtype as input, residing in GPU memory

Usage Examples

Example: Resize After Decode in PyTorch Integration Pipeline

from nvidia.dali import pipeline_def, fn, types

@pipeline_def
def image_pipe(img_hw=(320, 200)):
    encoded_images = fn.external_source(name="images", no_copy=True)

    decoded = fn.decoders.image(
        encoded_images,
        device="mixed",
        output_type=types.RGB,
        use_fast_idct=False,
        jpeg_fancy_upsampling=True,
    )

    images = fn.resize(
        decoded,
        size=img_hw,
        interp_type=types.INTERP_LINEAR,
        antialias=False,
    )

    return images

pipe = image_pipe(batch_size=8, num_threads=4, device_id=0)
pipe.build()

Example: Resize with Custom Dimensions

from nvidia.dali import pipeline_def, fn, types

@pipeline_def(batch_size=16, num_threads=4, device_id=0)
def resize_pipe():
    data = fn.external_source(name="input_images", no_copy=True)
    decoded = fn.decoders.image(data, device="mixed", output_type=types.RGB)
    # Resize to 224x224 for ImageNet-style models
    resized = fn.resize(
        decoded,
        size=(224, 224),
        interp_type=types.INTERP_LINEAR,
        antialias=True,
    )
    return resized

Related Pages

Implements Principle

• Principle:NVIDIA_DALI_Image_Resize

Requires Environment

• Environment:NVIDIA_DALI_CUDA_GPU_Environment